US 12,456,052 B2
Systems and methods for facilitating verifiability of machine learning model unlearning
Vijay Dialani, Fremont, CA (US); Bhaskar Krishnamachari, Los Angeles, CA (US); and Rajarshi Gupta, Los Altos, CA (US)
Assigned to Coinbase, Inc., Oakland, CA (US)
Filed by Coinbase, Inc., Oakland, CA (US)
Filed on Dec. 11, 2023, as Appl. No. 18/535,767.
Prior Publication US 2025/0190784 A1, Jun. 12, 2025
Int. Cl. G06N 3/08 (2023.01); G06F 16/22 (2019.01); G06F 16/27 (2019.01); G06N 3/0475 (2023.01); H04L 9/00 (2022.01); H04L 9/32 (2006.01)
CPC G06N 3/08 (2013.01) [G06F 16/2255 (2019.01); G06F 16/27 (2019.01); G06N 3/0475 (2023.01); H04L 9/3218 (2013.01); H04L 9/50 (2022.05)] 20 Claims
OG exemplary drawing
 
1. A system for facilitating generative model unlearning verifiability, the system comprising:
one or more processors and memories storing instructions that, when executed by the one or more processors, cause operations comprising:
executing a generative machine learning model trained on a training dataset, the training dataset comprising a first dataset that is subsequently indicated by a removal request to be removed from the generative machine learning model;
in connection with the removal request, generating token subsets corresponding to data subsets of the first dataset, and performing an unlearning process on the generative machine learning model that involves multiple unlearning stages at which different ones of the token subsets are used to respectively generate intermediate model instances of the generative machine learning model, the multiple unlearning stages comprising a first unlearning stage and a second unlearning stage that depends from the first unlearning stage;
in connection with the unlearning process, generating model instance hash-based values corresponding to the intermediate model instances, the model instance hash-based values comprising (i) a first hash-based value corresponding to a first intermediate model instance derived from a starting model instance of the generative machine learning model at the first unlearning stage of the unlearning process and (ii) a second hash-based value corresponding to a second intermediate model instance derived from the first intermediate model instance at the second unlearning stage of the unlearning process; and
causing storage of the model instance hash-based values at database locations of a distributed database configured with a consensus mechanism for committing to a synchronized state of the distributed database across nodes of the distributed database, the storage of the model instance hash-based values at the database locations enabling a verification system to verify performance of the unlearning process.